OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking
收藏Hugging Face2026-02-03 更新2026-02-07 收录
下载链接：
https://hf-mirror.com/datasets/OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking
下载链接
链接失效反馈官方服务：
资源简介：
MMFineReason-SFT-123K是一个多模态推理数据集，它是从MMFineReason-1.8M数据集中筛选出的最具挑战性的7%样本。这些样本的特点是Qwen3-VL-4B-Thinking模型在所有4次推理尝试中均失败。数据集包含123,000个样本，旨在通过更少的数据实现更高效的训练，同时保持与完整数据集相当的性能。数据集涵盖了视觉问答、问题回答和文本生成等任务，适用于多模态、推理、数学、科学等领域的研究。
MMFineReason-SFT-123K is a difficulty-filtered subset of MMFineReason-1.8M, containing only the hardest 7% of samples where Qwen3-VL-4B-Thinking consistently fails (pass rate = 0). It includes 123K challenging samples and aims to achieve comparable performance to the full 1.8M dataset with only 7% of the data. The dataset covers tasks such as visual-question-answering, question-answering, and text-generation, and is suitable for research in multimodal, reasoning, mathematics, science, and other STEM fields.
提供机构：
OpenDataArena
